Quality assessment of cross-topic article features based on improved CTS model

Assessing article quality is an important task in the field of education. To improve the semantic understanding of the assessment model and improve the generalizability of the article quality assessment system at the same time, an improved cross-topic article features quality assessment method based on CTS (Cross-prompt Trait Scorer) is proposed. In the shared low-level layer part, a new depth-separable convolution with residual structure is used to extract more semantic feature information; in the feature enhancement part, multi-headed attention is added to strengthen the effective features and improve the learning ability of lexicality; finally, a Log-Cosh loss function is used to make the model more robust to outliers and improve the model's QWK (Quadratic Weighted Kappa) evaluation metric of the model. The improved method experimented on the ASAP (Automated Student Assessment Prize) dataset, and the average QWK score reached 0.571 under different topics and 0.566 under different features. The experimental results showed that the improved model obtained superior results in the cross-topic article feature quality assessment task compared to the CTS and PAES (Prompt Agnostic Essay Scorer) models.